Billing

Updated 2026-11-11

Billing is prepaid: top up first, then pay for what you actually use. Each instance settles in a single currency, and prices, balances and bills are all shown in it.

Where your balance comes from

• Top-ups: place an order on the console Top-up page and pay online. Unpaid orders close after 60 minutes. A platform service fee is charged on top of each top-up at the rate shown on the pricing page, and you see it before you pay; it is not added to your balance.
• Redemption codes: credited as paid balance or gift credit, depending on the code.
• Gift credit: from trials and campaigns. It has an expiry date and is cleared when it expires.

Charges use the gift credit that expires soonest first, then paid balance. Gift credit only works on models marked "accepts gift credit" in the catalog, and it cannot be refunded or invoiced.

How prices work

Each model has its own price, published on the pricing page and the model page:

• Each call is billed at the price version in effect when the attempt that serves it starts; later price changes do not affect that call. Price cuts apply immediately; increases are announced in advance, and users who called that offering in the last 30 days are notified.
• If a model has no separate cache price, cached tokens are billed at the input price. Reasoning tokens are already part of output tokens and are not billed twice.
• The call amount is the sum of rate × quantity over all items, rounded half up to the smallest currency unit (0.000001 USD).

What is not charged

• Failed or canceled calls are free, and failed attempts are not paid to providers either; see Attempts and settlement.
• Calls whose result never arrives and stays unknown are free.
• If a stream is cut off midway, you pay for the usage already produced. When exact usage is unavailable, the charge is estimated and marked "estimated" in your billing lines.

Low balance and arrears

• Before each call the platform estimates the cost at each candidate offering's own price (estimated input plus the maximum output plus any per-request fee) and holds the highest one. Offerings your balance cannot pay for are not tried.
• If the request sets no maximum output, the platform works out how many output tokens your balance pays for and fills that in; offerings left with fewer than 1,000 are skipped. When no offering is affordable the call returns insufficient_balance (HTTP 402), and the message says how many output tokens your balance covers.
• If the actual cost exceeds the hold, the difference is charged as well, which can push paid balance below zero. The account is then in arrears and new calls return insufficient_balance (HTTP 402). Gift credit does not offset arrears; top up to zero or above to continue.
• Media tasks such as videos hold the estimated cost when created, settle at the actual cost on success, and release the hold on failure or cancellation.

Refunds and invoices

• Only the unspent paid amount of an order can be refunded; the service fee paid at top-up is not. The request asks for your login password; see Refunds.
• Invoices are issued for paid top-up orders only. Receipts and invoices list the top-up and the service fee on separate lines; gift credit is not invoiced.